Classification estimating system and classification estimating program

Abstract

A classification estimating system can include an input element group; an intermediate element group into which are input values of first intermediate variables for which a first multiple desensitization, including an accumulation of values of each input element of a first input element group and each value of a first output sensitivity and each value of a first multiplex output sensitivity, has been carried out and calculated; an output element group into which is input a value of an output variable calculated based on a value of each intermediate element of a first intermediate element group and a value of each connection weight; a classification information estimating module for estimating classification information based on pre-stored correspondence relationship information and a value of the output variable; and a classification information display.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS[0001]This application is a Continuation of International Application Number PCT / JP2010 / 004585, filed Jul. 14, 2010, which claims priority to Japanese Application No. 2009-167222, filed Jul. 15, 2009, the entire contents of both which are hereby incorporated by reference.BACKGROUND OF THE INVENTIONS[0002]1. Technical Field[0003]The present inventions relate to a classification estimating system and classification estimating program through a selective desensitization neural network for estimating classification information including an operational decision of a subject.[0004]2. Technical Background[0005]In recent years, research and development on mechanical control using physiological information (biosignals) of human beings are overwhelmingly carried out. Said mechanical control is designed not only to provide assistance in power assistance to prosthetic arms and limbs, or self-help devices through assistive techniques and adaptive devices fo...

AI Technical Summary

Benefits of technology

[0066]According to the invention described in claims 1 and 10, it is possible to estimate the classification information by way of inputting each value of the input variables based on the first and second information into the selective desensitization neutral network that enables multiple desensitization, and classification information is estimated through said selective desensitization neutral network with high training ability (generalization ability); therefore, compared with the classification information estimation where the selective desensitization neutral network that enables multiple desensitization is not applied, classification information estimation can be carried out with less learning requirement.

[0067]According to the invention described in claim 2, motion classification of a subject is carried out through an mutual modification model of the selective desensitization neutral network that enables multiple desensitization, therefore the values of two input variables are modified mutually (Multiple Mutually Desensitized) to make it possible for the calculation of two intermediate variables; which hence results in an improved training ability for the selective desensitization neutral network that enables successful multiple desensitization with high classification rate in estimating classification information compared with the case when the values of two input variables are modified through product-type modification (where only one value being desensitized).

[0069]According to the invention described in claim 4, compared with each calculated value of the elements of multiple element group based on the first or second information, each calculated value of the element of multiple element group based on the first and second information tends to be with high possibility, different from each value of the element of multiple element group based on the first or second information; the values of desensitized input elements tend to be difficult to deviate, which results in a high training ability for the selective desensitization neutral network that enables multiple desensitization and successful estimation of classification with comparatively high classification rate.

[0071]According to the invention described in claim 6, the operational decision of a subject can be estimated by inputting the values of the input variables based on the calculated average integrated value of the surface EMG signal per shift time into the selective desensitization neutral network. Compared with the case when inputting the values of the input variables directly based on the value of the surface EMG signal and its integrated value as biosignals, the interference caused by noise of surface EMG signal can be reduced. Moreover, according to the invention described in claim 6, the change of surface EMG signals accompanying the motions of a subject is detected 30 [ms]˜100 [ms] earlier than the muscle contraction accompanying the motions of a subject, hence, shift time can be decided to be shorter than the pre-set 30 [ms] resulting in a successful real time motion classification of a subject.

Problems solved by technology

Thus, with said techniques described in non-patent literatures 1 and 2, there exists a problem that motion classification is impossible when different velocity models are concerned.

In other words, the techniques described in said non-patent literatures 1 and 2 are applied to perform motion classification based on statistic approach to result in the necessity of a plurality of training data samples, which furthermore causes it burdensome to a subject hence to be deemed as impractical.

Accordingly, it is difficult to carry out real-time classification for complicated motions by means of the techniques described in said non-patent literatures 1 and 2.

Consequently, with techniques described in said non-patent literatures 1 and 2, it is difficult to carry out real-time classification for a plurality of motions together with motion velocities at high classification rate with little training data (training, sample data).

Images